Dev Kapadia 23′
Figure 1:
Example of a two-dimensional Lewis structure for water (dihydrogen monoxide). The dots indicate electrons, the letters indicate atoms, and the lines represent covalent bonds. Quantitative structure-activity or property relationship analyses indicate that all chemical information about a reaction can be derived from analysis of the Lewis structures of the components
Source: Wiki Commons
When we imagine the future, machine learning – an application of artificial intelligence (AI) algorithms – is almost always a feature of our imagination. But machine learning is already a feature of present day, used in medical diagnostics, language translation, and even autonomous cars. Recently, the power of machine learning has also captured the attention of chemists.1
Chemical reactions are extremely complex. Scientists need to consider whether or not the reactions will proceed at all and at what rate, both dependent on the identity of the reactants, their orientation relative to other molecules, and the amount of energy they possess. Even experienced chemists have trouble predicting their outcome with high accuracy due to the large amounts of information that need to be analyzed to make these predictions. However, while humans fall short in analyzing complex systems, machines excel. To improve our understanding of chemical reactions, chemists and computer scientists at the University of Münster in Germany produced an algorithm that uses the structural representations of reactants to predict information about chemical reactions.1
Previously, studies have demonstrated that the two-dimensional Lewis structure of molecules can be used to determine all information about chemical reactions [Figure 1]. The team at the University of Münster designed a program that could analyze these structures by making a graph that modeled the chemical environment, which resulted in a unique “fingerprint” for each molecular structure. While the researchers did not disclose the specifics of how the fingerprints were developed, they claimed that they were robust enough to uniquely represent each reaction component that is inputted into the algorithm. Once a large number of these fingerprints were uploaded, the algorithm was trained with these fingerprints and observational data so that the algorithm became robust enough to predict results for reactions not in the database. One unique aspect of the German team’s algorithm compared to others that are specific to reactions is that the German team’s algorithm can predict a reaction’s yield, amount of the products, and stereoselectivity, a characteristic where a single reactant forms a mixed amount of the stereoisomers.2,3
However, the algorithm does have shortcomings. For one, the data set that the team used for training is only designed to output relative yields, as opposed to absolute yields. Relative yields tell researchers how much product results as a percentage of how much is expected while absolute yield will give the amount of pure, dry product yielded from the reaction. To get the exact numerical, absolute yields, Frederik Sandfort, lead author of the paper detailing the algorithm, claims that “calibrations have to be created. However, due to the high effort involved, this is rarely done in reality.”1
The team designed this algorithm to analyze molecular structures and predict the outcome of new chemical reactions. This technology might be seen as a replacement for the jobs of synthetic chemists, tasked with the creation of new molecules by manipulating chemical reactions. But rather than see this as a replacement for the work of synthetic chemists, the German team see this more as a support of synthetic chemist’s work. The team hopes that their work on reaction prediction will help promote the future use of AI for synthetic chemistry.1
References:
[1] Predicting reaction results: Machines learn chemistry: Chemists and computer scientists develop artificial intelligence. (2020, March 17). ScienceDaily. Retrieved March 22, 2020 from www.sciencedaily.com/releases/2020/03/200317130711.htm
[2] Sandfort, F., Strieth-Kalthoff, F., Kühnemund, M., Beecks, C., & Glorius, F. (2020). A Structure-Based Platform for Predicting Chemical Reactivity. Chem. doi: 10.1016/j.chempr.2020.02.017
[3] The University of Manchester. (n.d.). Lecture 8: Stereoselective reactions. Retrieved March 22, 2020, from https://personalpages.manchester.ac.uk/staff/T.Wallace/20412tw2/chem20412_stereo_Lect8.htm